Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * SpanDSP - a series of DSP components for telephony
4 *
5 * echo.c - A line echo canceller. This code is being developed
6 * against and partially complies with G168.
7 *
8 * Written by Steve Underwood <steveu@coppice.org>
9 * and David Rowe <david_at_rowetel_dot_com>
10 *
11 * Copyright (C) 2001, 2003 Steve Underwood, 2007 David Rowe
12 *
13 * Based on a bit from here, a bit from there, eye of toad, ear of
14 * bat, 15 years of failed attempts by David and a few fried brain
15 * cells.
16 *
17 * All rights reserved.
18 */
19
20/*! \file */
21
22/* Implementation Notes
23 David Rowe
24 April 2007
25
26 This code started life as Steve's NLMS algorithm with a tap
27 rotation algorithm to handle divergence during double talk. I
28 added a Geigel Double Talk Detector (DTD) [2] and performed some
29 G168 tests. However I had trouble meeting the G168 requirements,
30 especially for double talk - there were always cases where my DTD
31 failed, for example where near end speech was under the 6dB
32 threshold required for declaring double talk.
33
34 So I tried a two path algorithm [1], which has so far given better
35 results. The original tap rotation/Geigel algorithm is available
36 in SVN http://svn.rowetel.com/software/oslec/tags/before_16bit.
37 It's probably possible to make it work if some one wants to put some
38 serious work into it.
39
40 At present no special treatment is provided for tones, which
41 generally cause NLMS algorithms to diverge. Initial runs of a
42 subset of the G168 tests for tones (e.g ./echo_test 6) show the
43 current algorithm is passing OK, which is kind of surprising. The
44 full set of tests needs to be performed to confirm this result.
45
46 One other interesting change is that I have managed to get the NLMS
47 code to work with 16 bit coefficients, rather than the original 32
48 bit coefficents. This reduces the MIPs and storage required.
49 I evaulated the 16 bit port using g168_tests.sh and listening tests
50 on 4 real-world samples.
51
52 I also attempted the implementation of a block based NLMS update
53 [2] but although this passes g168_tests.sh it didn't converge well
54 on the real-world samples. I have no idea why, perhaps a scaling
55 problem. The block based code is also available in SVN
56 http://svn.rowetel.com/software/oslec/tags/before_16bit. If this
57 code can be debugged, it will lead to further reduction in MIPS, as
58 the block update code maps nicely onto DSP instruction sets (it's a
59 dot product) compared to the current sample-by-sample update.
60
61 Steve also has some nice notes on echo cancellers in echo.h
62
63 References:
64
65 [1] Ochiai, Areseki, and Ogihara, "Echo Canceller with Two Echo
66 Path Models", IEEE Transactions on communications, COM-25,
67 No. 6, June
68 1977.
69 https://www.rowetel.com/images/echo/dual_path_paper.pdf
70
71 [2] The classic, very useful paper that tells you how to
72 actually build a real world echo canceller:
73 Messerschmitt, Hedberg, Cole, Haoui, Winship, "Digital Voice
74 Echo Canceller with a TMS320020,
75 https://www.rowetel.com/images/echo/spra129.pdf
76
77 [3] I have written a series of blog posts on this work, here is
78 Part 1: http://www.rowetel.com/blog/?p=18
79
80 [4] The source code http://svn.rowetel.com/software/oslec/
81
82 [5] A nice reference on LMS filters:
83 https://en.wikipedia.org/wiki/Least_mean_squares_filter
84
85 Credits:
86
87 Thanks to Steve Underwood, Jean-Marc Valin, and Ramakrishnan
88 Muthukrishnan for their suggestions and email discussions. Thanks
89 also to those people who collected echo samples for me such as
90 Mark, Pawel, and Pavel.
91*/
92
93#include <linux/kernel.h>
94#include <linux/module.h>
95#include <linux/slab.h>
96
97#include "echo.h"
98
99#define MIN_TX_POWER_FOR_ADAPTION 64
100#define MIN_RX_POWER_FOR_ADAPTION 64
101#define DTD_HANGOVER 600 /* 600 samples, or 75ms */
102#define DC_LOG2BETA 3 /* log2() of DC filter Beta */
103
104/* adapting coeffs using the traditional stochastic descent (N)LMS algorithm */
105
106static inline void lms_adapt_bg(struct oslec_state *ec, int clean, int shift)
107{
108 int i;
109
110 int offset1;
111 int offset2;
112 int factor;
113 int exp;
114
115 if (shift > 0)
116 factor = clean << shift;
117 else
118 factor = clean >> -shift;
119
120 /* Update the FIR taps */
121
122 offset2 = ec->curr_pos;
123 offset1 = ec->taps - offset2;
124
125 for (i = ec->taps - 1; i >= offset1; i--) {
126 exp = (ec->fir_state_bg.history[i - offset1] * factor);
127 ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15);
128 }
129 for (; i >= 0; i--) {
130 exp = (ec->fir_state_bg.history[i + offset2] * factor);
131 ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15);
132 }
133}
134
135static inline int top_bit(unsigned int bits)
136{
137 if (bits == 0)
138 return -1;
139 else
140 return (int)fls((int32_t) bits) - 1;
141}
142
143struct oslec_state *oslec_create(int len, int adaption_mode)
144{
145 struct oslec_state *ec;
146 int i;
147 const int16_t *history;
148
149 ec = kzalloc(sizeof(*ec), GFP_KERNEL);
150 if (!ec)
151 return NULL;
152
153 ec->taps = len;
154 ec->log2taps = top_bit(len);
155 ec->curr_pos = ec->taps - 1;
156
157 ec->fir_taps16[0] =
158 kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL);
159 if (!ec->fir_taps16[0])
160 goto error_oom_0;
161
162 ec->fir_taps16[1] =
163 kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL);
164 if (!ec->fir_taps16[1])
165 goto error_oom_1;
166
167 history = fir16_create(&ec->fir_state, ec->fir_taps16[0], ec->taps);
168 if (!history)
169 goto error_state;
170 history = fir16_create(&ec->fir_state_bg, ec->fir_taps16[1], ec->taps);
171 if (!history)
172 goto error_state_bg;
173
174 for (i = 0; i < 5; i++)
175 ec->xvtx[i] = ec->yvtx[i] = ec->xvrx[i] = ec->yvrx[i] = 0;
176
177 ec->cng_level = 1000;
178 oslec_adaption_mode(ec, adaption_mode);
179
180 ec->snapshot = kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL);
181 if (!ec->snapshot)
182 goto error_snap;
183
184 ec->cond_met = 0;
185 ec->pstates = 0;
186 ec->ltxacc = ec->lrxacc = ec->lcleanacc = ec->lclean_bgacc = 0;
187 ec->ltx = ec->lrx = ec->lclean = ec->lclean_bg = 0;
188 ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0;
189 ec->lbgn = ec->lbgn_acc = 0;
190 ec->lbgn_upper = 200;
191 ec->lbgn_upper_acc = ec->lbgn_upper << 13;
192
193 return ec;
194
195error_snap:
196 fir16_free(&ec->fir_state_bg);
197error_state_bg:
198 fir16_free(&ec->fir_state);
199error_state:
200 kfree(ec->fir_taps16[1]);
201error_oom_1:
202 kfree(ec->fir_taps16[0]);
203error_oom_0:
204 kfree(ec);
205 return NULL;
206}
207EXPORT_SYMBOL_GPL(oslec_create);
208
209void oslec_free(struct oslec_state *ec)
210{
211 int i;
212
213 fir16_free(&ec->fir_state);
214 fir16_free(&ec->fir_state_bg);
215 for (i = 0; i < 2; i++)
216 kfree(ec->fir_taps16[i]);
217 kfree(ec->snapshot);
218 kfree(ec);
219}
220EXPORT_SYMBOL_GPL(oslec_free);
221
222void oslec_adaption_mode(struct oslec_state *ec, int adaption_mode)
223{
224 ec->adaption_mode = adaption_mode;
225}
226EXPORT_SYMBOL_GPL(oslec_adaption_mode);
227
228void oslec_flush(struct oslec_state *ec)
229{
230 int i;
231
232 ec->ltxacc = ec->lrxacc = ec->lcleanacc = ec->lclean_bgacc = 0;
233 ec->ltx = ec->lrx = ec->lclean = ec->lclean_bg = 0;
234 ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0;
235
236 ec->lbgn = ec->lbgn_acc = 0;
237 ec->lbgn_upper = 200;
238 ec->lbgn_upper_acc = ec->lbgn_upper << 13;
239
240 ec->nonupdate_dwell = 0;
241
242 fir16_flush(&ec->fir_state);
243 fir16_flush(&ec->fir_state_bg);
244 ec->fir_state.curr_pos = ec->taps - 1;
245 ec->fir_state_bg.curr_pos = ec->taps - 1;
246 for (i = 0; i < 2; i++)
247 memset(ec->fir_taps16[i], 0, ec->taps * sizeof(int16_t));
248
249 ec->curr_pos = ec->taps - 1;
250 ec->pstates = 0;
251}
252EXPORT_SYMBOL_GPL(oslec_flush);
253
254void oslec_snapshot(struct oslec_state *ec)
255{
256 memcpy(ec->snapshot, ec->fir_taps16[0], ec->taps * sizeof(int16_t));
257}
258EXPORT_SYMBOL_GPL(oslec_snapshot);
259
260/* Dual Path Echo Canceller */
261
262int16_t oslec_update(struct oslec_state *ec, int16_t tx, int16_t rx)
263{
264 int32_t echo_value;
265 int clean_bg;
266 int tmp;
267 int tmp1;
268
269 /*
270 * Input scaling was found be required to prevent problems when tx
271 * starts clipping. Another possible way to handle this would be the
272 * filter coefficent scaling.
273 */
274
275 ec->tx = tx;
276 ec->rx = rx;
277 tx >>= 1;
278 rx >>= 1;
279
280 /*
281 * Filter DC, 3dB point is 160Hz (I think), note 32 bit precision
282 * required otherwise values do not track down to 0. Zero at DC, Pole
283 * at (1-Beta) on real axis. Some chip sets (like Si labs) don't
284 * need this, but something like a $10 X100P card does. Any DC really
285 * slows down convergence.
286 *
287 * Note: removes some low frequency from the signal, this reduces the
288 * speech quality when listening to samples through headphones but may
289 * not be obvious through a telephone handset.
290 *
291 * Note that the 3dB frequency in radians is approx Beta, e.g. for Beta
292 * = 2^(-3) = 0.125, 3dB freq is 0.125 rads = 159Hz.
293 */
294
295 if (ec->adaption_mode & ECHO_CAN_USE_RX_HPF) {
296 tmp = rx << 15;
297
298 /*
299 * Make sure the gain of the HPF is 1.0. This can still
300 * saturate a little under impulse conditions, and it might
301 * roll to 32768 and need clipping on sustained peak level
302 * signals. However, the scale of such clipping is small, and
303 * the error due to any saturation should not markedly affect
304 * the downstream processing.
305 */
306 tmp -= (tmp >> 4);
307
308 ec->rx_1 += -(ec->rx_1 >> DC_LOG2BETA) + tmp - ec->rx_2;
309
310 /*
311 * hard limit filter to prevent clipping. Note that at this
312 * stage rx should be limited to +/- 16383 due to right shift
313 * above
314 */
315 tmp1 = ec->rx_1 >> 15;
316 if (tmp1 > 16383)
317 tmp1 = 16383;
318 if (tmp1 < -16383)
319 tmp1 = -16383;
320 rx = tmp1;
321 ec->rx_2 = tmp;
322 }
323
324 /* Block average of power in the filter states. Used for
325 adaption power calculation. */
326
327 {
328 int new, old;
329
330 /* efficient "out with the old and in with the new" algorithm so
331 we don't have to recalculate over the whole block of
332 samples. */
333 new = (int)tx * (int)tx;
334 old = (int)ec->fir_state.history[ec->fir_state.curr_pos] *
335 (int)ec->fir_state.history[ec->fir_state.curr_pos];
336 ec->pstates +=
337 ((new - old) + (1 << (ec->log2taps - 1))) >> ec->log2taps;
338 if (ec->pstates < 0)
339 ec->pstates = 0;
340 }
341
342 /* Calculate short term average levels using simple single pole IIRs */
343
344 ec->ltxacc += abs(tx) - ec->ltx;
345 ec->ltx = (ec->ltxacc + (1 << 4)) >> 5;
346 ec->lrxacc += abs(rx) - ec->lrx;
347 ec->lrx = (ec->lrxacc + (1 << 4)) >> 5;
348
349 /* Foreground filter */
350
351 ec->fir_state.coeffs = ec->fir_taps16[0];
352 echo_value = fir16(&ec->fir_state, tx);
353 ec->clean = rx - echo_value;
354 ec->lcleanacc += abs(ec->clean) - ec->lclean;
355 ec->lclean = (ec->lcleanacc + (1 << 4)) >> 5;
356
357 /* Background filter */
358
359 echo_value = fir16(&ec->fir_state_bg, tx);
360 clean_bg = rx - echo_value;
361 ec->lclean_bgacc += abs(clean_bg) - ec->lclean_bg;
362 ec->lclean_bg = (ec->lclean_bgacc + (1 << 4)) >> 5;
363
364 /* Background Filter adaption */
365
366 /* Almost always adap bg filter, just simple DT and energy
367 detection to minimise adaption in cases of strong double talk.
368 However this is not critical for the dual path algorithm.
369 */
370 ec->factor = 0;
371 ec->shift = 0;
372 if (!ec->nonupdate_dwell) {
373 int p, logp, shift;
374
375 /* Determine:
376
377 f = Beta * clean_bg_rx/P ------ (1)
378
379 where P is the total power in the filter states.
380
381 The Boffins have shown that if we obey (1) we converge
382 quickly and avoid instability.
383
384 The correct factor f must be in Q30, as this is the fixed
385 point format required by the lms_adapt_bg() function,
386 therefore the scaled version of (1) is:
387
388 (2^30) * f = (2^30) * Beta * clean_bg_rx/P
389 factor = (2^30) * Beta * clean_bg_rx/P ----- (2)
390
391 We have chosen Beta = 0.25 by experiment, so:
392
393 factor = (2^30) * (2^-2) * clean_bg_rx/P
394
395 (30 - 2 - log2(P))
396 factor = clean_bg_rx 2 ----- (3)
397
398 To avoid a divide we approximate log2(P) as top_bit(P),
399 which returns the position of the highest non-zero bit in
400 P. This approximation introduces an error as large as a
401 factor of 2, but the algorithm seems to handle it OK.
402
403 Come to think of it a divide may not be a big deal on a
404 modern DSP, so its probably worth checking out the cycles
405 for a divide versus a top_bit() implementation.
406 */
407
408 p = MIN_TX_POWER_FOR_ADAPTION + ec->pstates;
409 logp = top_bit(p) + ec->log2taps;
410 shift = 30 - 2 - logp;
411 ec->shift = shift;
412
413 lms_adapt_bg(ec, clean_bg, shift);
414 }
415
416 /* very simple DTD to make sure we dont try and adapt with strong
417 near end speech */
418
419 ec->adapt = 0;
420 if ((ec->lrx > MIN_RX_POWER_FOR_ADAPTION) && (ec->lrx > ec->ltx))
421 ec->nonupdate_dwell = DTD_HANGOVER;
422 if (ec->nonupdate_dwell)
423 ec->nonupdate_dwell--;
424
425 /* Transfer logic */
426
427 /* These conditions are from the dual path paper [1], I messed with
428 them a bit to improve performance. */
429
430 if ((ec->adaption_mode & ECHO_CAN_USE_ADAPTION) &&
431 (ec->nonupdate_dwell == 0) &&
432 /* (ec->Lclean_bg < 0.875*ec->Lclean) */
433 (8 * ec->lclean_bg < 7 * ec->lclean) &&
434 /* (ec->Lclean_bg < 0.125*ec->Ltx) */
435 (8 * ec->lclean_bg < ec->ltx)) {
436 if (ec->cond_met == 6) {
437 /*
438 * BG filter has had better results for 6 consecutive
439 * samples
440 */
441 ec->adapt = 1;
442 memcpy(ec->fir_taps16[0], ec->fir_taps16[1],
443 ec->taps * sizeof(int16_t));
444 } else
445 ec->cond_met++;
446 } else
447 ec->cond_met = 0;
448
449 /* Non-Linear Processing */
450
451 ec->clean_nlp = ec->clean;
452 if (ec->adaption_mode & ECHO_CAN_USE_NLP) {
453 /*
454 * Non-linear processor - a fancy way to say "zap small
455 * signals, to avoid residual echo due to (uLaw/ALaw)
456 * non-linearity in the channel.".
457 */
458
459 if ((16 * ec->lclean < ec->ltx)) {
460 /*
461 * Our e/c has improved echo by at least 24 dB (each
462 * factor of 2 is 6dB, so 2*2*2*2=16 is the same as
463 * 6+6+6+6=24dB)
464 */
465 if (ec->adaption_mode & ECHO_CAN_USE_CNG) {
466 ec->cng_level = ec->lbgn;
467
468 /*
469 * Very elementary comfort noise generation.
470 * Just random numbers rolled off very vaguely
471 * Hoth-like. DR: This noise doesn't sound
472 * quite right to me - I suspect there are some
473 * overflow issues in the filtering as it's too
474 * "crackly".
475 * TODO: debug this, maybe just play noise at
476 * high level or look at spectrum.
477 */
478
479 ec->cng_rndnum =
480 1664525U * ec->cng_rndnum + 1013904223U;
481 ec->cng_filter =
482 ((ec->cng_rndnum & 0xFFFF) - 32768 +
483 5 * ec->cng_filter) >> 3;
484 ec->clean_nlp =
485 (ec->cng_filter * ec->cng_level * 8) >> 14;
486
487 } else if (ec->adaption_mode & ECHO_CAN_USE_CLIP) {
488 /* This sounds much better than CNG */
489 if (ec->clean_nlp > ec->lbgn)
490 ec->clean_nlp = ec->lbgn;
491 if (ec->clean_nlp < -ec->lbgn)
492 ec->clean_nlp = -ec->lbgn;
493 } else {
494 /*
495 * just mute the residual, doesn't sound very
496 * good, used mainly in G168 tests
497 */
498 ec->clean_nlp = 0;
499 }
500 } else {
501 /*
502 * Background noise estimator. I tried a few
503 * algorithms here without much luck. This very simple
504 * one seems to work best, we just average the level
505 * using a slow (1 sec time const) filter if the
506 * current level is less than a (experimentally
507 * derived) constant. This means we dont include high
508 * level signals like near end speech. When combined
509 * with CNG or especially CLIP seems to work OK.
510 */
511 if (ec->lclean < 40) {
512 ec->lbgn_acc += abs(ec->clean) - ec->lbgn;
513 ec->lbgn = (ec->lbgn_acc + (1 << 11)) >> 12;
514 }
515 }
516 }
517
518 /* Roll around the taps buffer */
519 if (ec->curr_pos <= 0)
520 ec->curr_pos = ec->taps;
521 ec->curr_pos--;
522
523 if (ec->adaption_mode & ECHO_CAN_DISABLE)
524 ec->clean_nlp = rx;
525
526 /* Output scaled back up again to match input scaling */
527
528 return (int16_t) ec->clean_nlp << 1;
529}
530EXPORT_SYMBOL_GPL(oslec_update);
531
532/* This function is separated from the echo canceller is it is usually called
533 as part of the tx process. See rx HP (DC blocking) filter above, it's
534 the same design.
535
536 Some soft phones send speech signals with a lot of low frequency
537 energy, e.g. down to 20Hz. This can make the hybrid non-linear
538 which causes the echo canceller to fall over. This filter can help
539 by removing any low frequency before it gets to the tx port of the
540 hybrid.
541
542 It can also help by removing and DC in the tx signal. DC is bad
543 for LMS algorithms.
544
545 This is one of the classic DC removal filters, adjusted to provide
546 sufficient bass rolloff to meet the above requirement to protect hybrids
547 from things that upset them. The difference between successive samples
548 produces a lousy HPF, and then a suitably placed pole flattens things out.
549 The final result is a nicely rolled off bass end. The filtering is
550 implemented with extended fractional precision, which noise shapes things,
551 giving very clean DC removal.
552*/
553
554int16_t oslec_hpf_tx(struct oslec_state *ec, int16_t tx)
555{
556 int tmp;
557 int tmp1;
558
559 if (ec->adaption_mode & ECHO_CAN_USE_TX_HPF) {
560 tmp = tx << 15;
561
562 /*
563 * Make sure the gain of the HPF is 1.0. The first can still
564 * saturate a little under impulse conditions, and it might
565 * roll to 32768 and need clipping on sustained peak level
566 * signals. However, the scale of such clipping is small, and
567 * the error due to any saturation should not markedly affect
568 * the downstream processing.
569 */
570 tmp -= (tmp >> 4);
571
572 ec->tx_1 += -(ec->tx_1 >> DC_LOG2BETA) + tmp - ec->tx_2;
573 tmp1 = ec->tx_1 >> 15;
574 if (tmp1 > 32767)
575 tmp1 = 32767;
576 if (tmp1 < -32767)
577 tmp1 = -32767;
578 tx = tmp1;
579 ec->tx_2 = tmp;
580 }
581
582 return tx;
583}
584EXPORT_SYMBOL_GPL(oslec_hpf_tx);
585
586MODULE_LICENSE("GPL");
587MODULE_AUTHOR("David Rowe");
588MODULE_DESCRIPTION("Open Source Line Echo Canceller");
589MODULE_VERSION("0.3.0");
1/*
2 * SpanDSP - a series of DSP components for telephony
3 *
4 * echo.c - A line echo canceller. This code is being developed
5 * against and partially complies with G168.
6 *
7 * Written by Steve Underwood <steveu@coppice.org>
8 * and David Rowe <david_at_rowetel_dot_com>
9 *
10 * Copyright (C) 2001, 2003 Steve Underwood, 2007 David Rowe
11 *
12 * Based on a bit from here, a bit from there, eye of toad, ear of
13 * bat, 15 years of failed attempts by David and a few fried brain
14 * cells.
15 *
16 * All rights reserved.
17 *
18 * This program is free software; you can redistribute it and/or modify
19 * it under the terms of the GNU General Public License version 2, as
20 * published by the Free Software Foundation.
21 *
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
26 *
27 * You should have received a copy of the GNU General Public License
28 * along with this program; if not, write to the Free Software
29 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
30 */
31
32/*! \file */
33
34/* Implementation Notes
35 David Rowe
36 April 2007
37
38 This code started life as Steve's NLMS algorithm with a tap
39 rotation algorithm to handle divergence during double talk. I
40 added a Geigel Double Talk Detector (DTD) [2] and performed some
41 G168 tests. However I had trouble meeting the G168 requirements,
42 especially for double talk - there were always cases where my DTD
43 failed, for example where near end speech was under the 6dB
44 threshold required for declaring double talk.
45
46 So I tried a two path algorithm [1], which has so far given better
47 results. The original tap rotation/Geigel algorithm is available
48 in SVN http://svn.rowetel.com/software/oslec/tags/before_16bit.
49 It's probably possible to make it work if some one wants to put some
50 serious work into it.
51
52 At present no special treatment is provided for tones, which
53 generally cause NLMS algorithms to diverge. Initial runs of a
54 subset of the G168 tests for tones (e.g ./echo_test 6) show the
55 current algorithm is passing OK, which is kind of surprising. The
56 full set of tests needs to be performed to confirm this result.
57
58 One other interesting change is that I have managed to get the NLMS
59 code to work with 16 bit coefficients, rather than the original 32
60 bit coefficents. This reduces the MIPs and storage required.
61 I evaulated the 16 bit port using g168_tests.sh and listening tests
62 on 4 real-world samples.
63
64 I also attempted the implementation of a block based NLMS update
65 [2] but although this passes g168_tests.sh it didn't converge well
66 on the real-world samples. I have no idea why, perhaps a scaling
67 problem. The block based code is also available in SVN
68 http://svn.rowetel.com/software/oslec/tags/before_16bit. If this
69 code can be debugged, it will lead to further reduction in MIPS, as
70 the block update code maps nicely onto DSP instruction sets (it's a
71 dot product) compared to the current sample-by-sample update.
72
73 Steve also has some nice notes on echo cancellers in echo.h
74
75 References:
76
77 [1] Ochiai, Areseki, and Ogihara, "Echo Canceller with Two Echo
78 Path Models", IEEE Transactions on communications, COM-25,
79 No. 6, June
80 1977.
81 http://www.rowetel.com/images/echo/dual_path_paper.pdf
82
83 [2] The classic, very useful paper that tells you how to
84 actually build a real world echo canceller:
85 Messerschmitt, Hedberg, Cole, Haoui, Winship, "Digital Voice
86 Echo Canceller with a TMS320020,
87 http://www.rowetel.com/images/echo/spra129.pdf
88
89 [3] I have written a series of blog posts on this work, here is
90 Part 1: http://www.rowetel.com/blog/?p=18
91
92 [4] The source code http://svn.rowetel.com/software/oslec/
93
94 [5] A nice reference on LMS filters:
95 http://en.wikipedia.org/wiki/Least_mean_squares_filter
96
97 Credits:
98
99 Thanks to Steve Underwood, Jean-Marc Valin, and Ramakrishnan
100 Muthukrishnan for their suggestions and email discussions. Thanks
101 also to those people who collected echo samples for me such as
102 Mark, Pawel, and Pavel.
103*/
104
105#include <linux/kernel.h>
106#include <linux/module.h>
107#include <linux/slab.h>
108
109#include "echo.h"
110
111#define MIN_TX_POWER_FOR_ADAPTION 64
112#define MIN_RX_POWER_FOR_ADAPTION 64
113#define DTD_HANGOVER 600 /* 600 samples, or 75ms */
114#define DC_LOG2BETA 3 /* log2() of DC filter Beta */
115
116/* adapting coeffs using the traditional stochastic descent (N)LMS algorithm */
117
118#ifdef __bfin__
119static inline void lms_adapt_bg(struct oslec_state *ec, int clean, int shift)
120{
121 int i;
122 int offset1;
123 int offset2;
124 int factor;
125 int exp;
126 int16_t *phist;
127 int n;
128
129 if (shift > 0)
130 factor = clean << shift;
131 else
132 factor = clean >> -shift;
133
134 /* Update the FIR taps */
135
136 offset2 = ec->curr_pos;
137 offset1 = ec->taps - offset2;
138 phist = &ec->fir_state_bg.history[offset2];
139
140 /* st: and en: help us locate the assembler in echo.s */
141
142 /* asm("st:"); */
143 n = ec->taps;
144 for (i = 0; i < n; i++) {
145 exp = *phist++ * factor;
146 ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15);
147 }
148 /* asm("en:"); */
149
150 /* Note the asm for the inner loop above generated by Blackfin gcc
151 4.1.1 is pretty good (note even parallel instructions used):
152
153 R0 = W [P0++] (X);
154 R0 *= R2;
155 R0 = R0 + R3 (NS) ||
156 R1 = W [P1] (X) ||
157 nop;
158 R0 >>>= 15;
159 R0 = R0 + R1;
160 W [P1++] = R0;
161
162 A block based update algorithm would be much faster but the
163 above can't be improved on much. Every instruction saved in
164 the loop above is 2 MIPs/ch! The for loop above is where the
165 Blackfin spends most of it's time - about 17 MIPs/ch measured
166 with speedtest.c with 256 taps (32ms). Write-back and
167 Write-through cache gave about the same performance.
168 */
169}
170
171/*
172 IDEAS for further optimisation of lms_adapt_bg():
173
174 1/ The rounding is quite costly. Could we keep as 32 bit coeffs
175 then make filter pluck the MS 16-bits of the coeffs when filtering?
176 However this would lower potential optimisation of filter, as I
177 think the dual-MAC architecture requires packed 16 bit coeffs.
178
179 2/ Block based update would be more efficient, as per comments above,
180 could use dual MAC architecture.
181
182 3/ Look for same sample Blackfin LMS code, see if we can get dual-MAC
183 packing.
184
185 4/ Execute the whole e/c in a block of say 20ms rather than sample
186 by sample. Processing a few samples every ms is inefficient.
187*/
188
189#else
190static inline void lms_adapt_bg(struct oslec_state *ec, int clean, int shift)
191{
192 int i;
193
194 int offset1;
195 int offset2;
196 int factor;
197 int exp;
198
199 if (shift > 0)
200 factor = clean << shift;
201 else
202 factor = clean >> -shift;
203
204 /* Update the FIR taps */
205
206 offset2 = ec->curr_pos;
207 offset1 = ec->taps - offset2;
208
209 for (i = ec->taps - 1; i >= offset1; i--) {
210 exp = (ec->fir_state_bg.history[i - offset1] * factor);
211 ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15);
212 }
213 for (; i >= 0; i--) {
214 exp = (ec->fir_state_bg.history[i + offset2] * factor);
215 ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15);
216 }
217}
218#endif
219
220static inline int top_bit(unsigned int bits)
221{
222 if (bits == 0)
223 return -1;
224 else
225 return (int)fls((int32_t) bits) - 1;
226}
227
228struct oslec_state *oslec_create(int len, int adaption_mode)
229{
230 struct oslec_state *ec;
231 int i;
232 const int16_t *history;
233
234 ec = kzalloc(sizeof(*ec), GFP_KERNEL);
235 if (!ec)
236 return NULL;
237
238 ec->taps = len;
239 ec->log2taps = top_bit(len);
240 ec->curr_pos = ec->taps - 1;
241
242 ec->fir_taps16[0] =
243 kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL);
244 if (!ec->fir_taps16[0])
245 goto error_oom_0;
246
247 ec->fir_taps16[1] =
248 kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL);
249 if (!ec->fir_taps16[1])
250 goto error_oom_1;
251
252 history = fir16_create(&ec->fir_state, ec->fir_taps16[0], ec->taps);
253 if (!history)
254 goto error_state;
255 history = fir16_create(&ec->fir_state_bg, ec->fir_taps16[1], ec->taps);
256 if (!history)
257 goto error_state_bg;
258
259 for (i = 0; i < 5; i++)
260 ec->xvtx[i] = ec->yvtx[i] = ec->xvrx[i] = ec->yvrx[i] = 0;
261
262 ec->cng_level = 1000;
263 oslec_adaption_mode(ec, adaption_mode);
264
265 ec->snapshot = kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL);
266 if (!ec->snapshot)
267 goto error_snap;
268
269 ec->cond_met = 0;
270 ec->pstates = 0;
271 ec->ltxacc = ec->lrxacc = ec->lcleanacc = ec->lclean_bgacc = 0;
272 ec->ltx = ec->lrx = ec->lclean = ec->lclean_bg = 0;
273 ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0;
274 ec->lbgn = ec->lbgn_acc = 0;
275 ec->lbgn_upper = 200;
276 ec->lbgn_upper_acc = ec->lbgn_upper << 13;
277
278 return ec;
279
280error_snap:
281 fir16_free(&ec->fir_state_bg);
282error_state_bg:
283 fir16_free(&ec->fir_state);
284error_state:
285 kfree(ec->fir_taps16[1]);
286error_oom_1:
287 kfree(ec->fir_taps16[0]);
288error_oom_0:
289 kfree(ec);
290 return NULL;
291}
292EXPORT_SYMBOL_GPL(oslec_create);
293
294void oslec_free(struct oslec_state *ec)
295{
296 int i;
297
298 fir16_free(&ec->fir_state);
299 fir16_free(&ec->fir_state_bg);
300 for (i = 0; i < 2; i++)
301 kfree(ec->fir_taps16[i]);
302 kfree(ec->snapshot);
303 kfree(ec);
304}
305EXPORT_SYMBOL_GPL(oslec_free);
306
307void oslec_adaption_mode(struct oslec_state *ec, int adaption_mode)
308{
309 ec->adaption_mode = adaption_mode;
310}
311EXPORT_SYMBOL_GPL(oslec_adaption_mode);
312
313void oslec_flush(struct oslec_state *ec)
314{
315 int i;
316
317 ec->ltxacc = ec->lrxacc = ec->lcleanacc = ec->lclean_bgacc = 0;
318 ec->ltx = ec->lrx = ec->lclean = ec->lclean_bg = 0;
319 ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0;
320
321 ec->lbgn = ec->lbgn_acc = 0;
322 ec->lbgn_upper = 200;
323 ec->lbgn_upper_acc = ec->lbgn_upper << 13;
324
325 ec->nonupdate_dwell = 0;
326
327 fir16_flush(&ec->fir_state);
328 fir16_flush(&ec->fir_state_bg);
329 ec->fir_state.curr_pos = ec->taps - 1;
330 ec->fir_state_bg.curr_pos = ec->taps - 1;
331 for (i = 0; i < 2; i++)
332 memset(ec->fir_taps16[i], 0, ec->taps * sizeof(int16_t));
333
334 ec->curr_pos = ec->taps - 1;
335 ec->pstates = 0;
336}
337EXPORT_SYMBOL_GPL(oslec_flush);
338
339void oslec_snapshot(struct oslec_state *ec)
340{
341 memcpy(ec->snapshot, ec->fir_taps16[0], ec->taps * sizeof(int16_t));
342}
343EXPORT_SYMBOL_GPL(oslec_snapshot);
344
345/* Dual Path Echo Canceller */
346
347int16_t oslec_update(struct oslec_state *ec, int16_t tx, int16_t rx)
348{
349 int32_t echo_value;
350 int clean_bg;
351 int tmp;
352 int tmp1;
353
354 /*
355 * Input scaling was found be required to prevent problems when tx
356 * starts clipping. Another possible way to handle this would be the
357 * filter coefficent scaling.
358 */
359
360 ec->tx = tx;
361 ec->rx = rx;
362 tx >>= 1;
363 rx >>= 1;
364
365 /*
366 * Filter DC, 3dB point is 160Hz (I think), note 32 bit precision
367 * required otherwise values do not track down to 0. Zero at DC, Pole
368 * at (1-Beta) on real axis. Some chip sets (like Si labs) don't
369 * need this, but something like a $10 X100P card does. Any DC really
370 * slows down convergence.
371 *
372 * Note: removes some low frequency from the signal, this reduces the
373 * speech quality when listening to samples through headphones but may
374 * not be obvious through a telephone handset.
375 *
376 * Note that the 3dB frequency in radians is approx Beta, e.g. for Beta
377 * = 2^(-3) = 0.125, 3dB freq is 0.125 rads = 159Hz.
378 */
379
380 if (ec->adaption_mode & ECHO_CAN_USE_RX_HPF) {
381 tmp = rx << 15;
382
383 /*
384 * Make sure the gain of the HPF is 1.0. This can still
385 * saturate a little under impulse conditions, and it might
386 * roll to 32768 and need clipping on sustained peak level
387 * signals. However, the scale of such clipping is small, and
388 * the error due to any saturation should not markedly affect
389 * the downstream processing.
390 */
391 tmp -= (tmp >> 4);
392
393 ec->rx_1 += -(ec->rx_1 >> DC_LOG2BETA) + tmp - ec->rx_2;
394
395 /*
396 * hard limit filter to prevent clipping. Note that at this
397 * stage rx should be limited to +/- 16383 due to right shift
398 * above
399 */
400 tmp1 = ec->rx_1 >> 15;
401 if (tmp1 > 16383)
402 tmp1 = 16383;
403 if (tmp1 < -16383)
404 tmp1 = -16383;
405 rx = tmp1;
406 ec->rx_2 = tmp;
407 }
408
409 /* Block average of power in the filter states. Used for
410 adaption power calculation. */
411
412 {
413 int new, old;
414
415 /* efficient "out with the old and in with the new" algorithm so
416 we don't have to recalculate over the whole block of
417 samples. */
418 new = (int)tx * (int)tx;
419 old = (int)ec->fir_state.history[ec->fir_state.curr_pos] *
420 (int)ec->fir_state.history[ec->fir_state.curr_pos];
421 ec->pstates +=
422 ((new - old) + (1 << (ec->log2taps - 1))) >> ec->log2taps;
423 if (ec->pstates < 0)
424 ec->pstates = 0;
425 }
426
427 /* Calculate short term average levels using simple single pole IIRs */
428
429 ec->ltxacc += abs(tx) - ec->ltx;
430 ec->ltx = (ec->ltxacc + (1 << 4)) >> 5;
431 ec->lrxacc += abs(rx) - ec->lrx;
432 ec->lrx = (ec->lrxacc + (1 << 4)) >> 5;
433
434 /* Foreground filter */
435
436 ec->fir_state.coeffs = ec->fir_taps16[0];
437 echo_value = fir16(&ec->fir_state, tx);
438 ec->clean = rx - echo_value;
439 ec->lcleanacc += abs(ec->clean) - ec->lclean;
440 ec->lclean = (ec->lcleanacc + (1 << 4)) >> 5;
441
442 /* Background filter */
443
444 echo_value = fir16(&ec->fir_state_bg, tx);
445 clean_bg = rx - echo_value;
446 ec->lclean_bgacc += abs(clean_bg) - ec->lclean_bg;
447 ec->lclean_bg = (ec->lclean_bgacc + (1 << 4)) >> 5;
448
449 /* Background Filter adaption */
450
451 /* Almost always adap bg filter, just simple DT and energy
452 detection to minimise adaption in cases of strong double talk.
453 However this is not critical for the dual path algorithm.
454 */
455 ec->factor = 0;
456 ec->shift = 0;
457 if ((ec->nonupdate_dwell == 0)) {
458 int p, logp, shift;
459
460 /* Determine:
461
462 f = Beta * clean_bg_rx/P ------ (1)
463
464 where P is the total power in the filter states.
465
466 The Boffins have shown that if we obey (1) we converge
467 quickly and avoid instability.
468
469 The correct factor f must be in Q30, as this is the fixed
470 point format required by the lms_adapt_bg() function,
471 therefore the scaled version of (1) is:
472
473 (2^30) * f = (2^30) * Beta * clean_bg_rx/P
474 factor = (2^30) * Beta * clean_bg_rx/P ----- (2)
475
476 We have chosen Beta = 0.25 by experiment, so:
477
478 factor = (2^30) * (2^-2) * clean_bg_rx/P
479
480 (30 - 2 - log2(P))
481 factor = clean_bg_rx 2 ----- (3)
482
483 To avoid a divide we approximate log2(P) as top_bit(P),
484 which returns the position of the highest non-zero bit in
485 P. This approximation introduces an error as large as a
486 factor of 2, but the algorithm seems to handle it OK.
487
488 Come to think of it a divide may not be a big deal on a
489 modern DSP, so its probably worth checking out the cycles
490 for a divide versus a top_bit() implementation.
491 */
492
493 p = MIN_TX_POWER_FOR_ADAPTION + ec->pstates;
494 logp = top_bit(p) + ec->log2taps;
495 shift = 30 - 2 - logp;
496 ec->shift = shift;
497
498 lms_adapt_bg(ec, clean_bg, shift);
499 }
500
501 /* very simple DTD to make sure we dont try and adapt with strong
502 near end speech */
503
504 ec->adapt = 0;
505 if ((ec->lrx > MIN_RX_POWER_FOR_ADAPTION) && (ec->lrx > ec->ltx))
506 ec->nonupdate_dwell = DTD_HANGOVER;
507 if (ec->nonupdate_dwell)
508 ec->nonupdate_dwell--;
509
510 /* Transfer logic */
511
512 /* These conditions are from the dual path paper [1], I messed with
513 them a bit to improve performance. */
514
515 if ((ec->adaption_mode & ECHO_CAN_USE_ADAPTION) &&
516 (ec->nonupdate_dwell == 0) &&
517 /* (ec->Lclean_bg < 0.875*ec->Lclean) */
518 (8 * ec->lclean_bg < 7 * ec->lclean) &&
519 /* (ec->Lclean_bg < 0.125*ec->Ltx) */
520 (8 * ec->lclean_bg < ec->ltx)) {
521 if (ec->cond_met == 6) {
522 /*
523 * BG filter has had better results for 6 consecutive
524 * samples
525 */
526 ec->adapt = 1;
527 memcpy(ec->fir_taps16[0], ec->fir_taps16[1],
528 ec->taps * sizeof(int16_t));
529 } else
530 ec->cond_met++;
531 } else
532 ec->cond_met = 0;
533
534 /* Non-Linear Processing */
535
536 ec->clean_nlp = ec->clean;
537 if (ec->adaption_mode & ECHO_CAN_USE_NLP) {
538 /*
539 * Non-linear processor - a fancy way to say "zap small
540 * signals, to avoid residual echo due to (uLaw/ALaw)
541 * non-linearity in the channel.".
542 */
543
544 if ((16 * ec->lclean < ec->ltx)) {
545 /*
546 * Our e/c has improved echo by at least 24 dB (each
547 * factor of 2 is 6dB, so 2*2*2*2=16 is the same as
548 * 6+6+6+6=24dB)
549 */
550 if (ec->adaption_mode & ECHO_CAN_USE_CNG) {
551 ec->cng_level = ec->lbgn;
552
553 /*
554 * Very elementary comfort noise generation.
555 * Just random numbers rolled off very vaguely
556 * Hoth-like. DR: This noise doesn't sound
557 * quite right to me - I suspect there are some
558 * overflow issues in the filtering as it's too
559 * "crackly".
560 * TODO: debug this, maybe just play noise at
561 * high level or look at spectrum.
562 */
563
564 ec->cng_rndnum =
565 1664525U * ec->cng_rndnum + 1013904223U;
566 ec->cng_filter =
567 ((ec->cng_rndnum & 0xFFFF) - 32768 +
568 5 * ec->cng_filter) >> 3;
569 ec->clean_nlp =
570 (ec->cng_filter * ec->cng_level * 8) >> 14;
571
572 } else if (ec->adaption_mode & ECHO_CAN_USE_CLIP) {
573 /* This sounds much better than CNG */
574 if (ec->clean_nlp > ec->lbgn)
575 ec->clean_nlp = ec->lbgn;
576 if (ec->clean_nlp < -ec->lbgn)
577 ec->clean_nlp = -ec->lbgn;
578 } else {
579 /*
580 * just mute the residual, doesn't sound very
581 * good, used mainly in G168 tests
582 */
583 ec->clean_nlp = 0;
584 }
585 } else {
586 /*
587 * Background noise estimator. I tried a few
588 * algorithms here without much luck. This very simple
589 * one seems to work best, we just average the level
590 * using a slow (1 sec time const) filter if the
591 * current level is less than a (experimentally
592 * derived) constant. This means we dont include high
593 * level signals like near end speech. When combined
594 * with CNG or especially CLIP seems to work OK.
595 */
596 if (ec->lclean < 40) {
597 ec->lbgn_acc += abs(ec->clean) - ec->lbgn;
598 ec->lbgn = (ec->lbgn_acc + (1 << 11)) >> 12;
599 }
600 }
601 }
602
603 /* Roll around the taps buffer */
604 if (ec->curr_pos <= 0)
605 ec->curr_pos = ec->taps;
606 ec->curr_pos--;
607
608 if (ec->adaption_mode & ECHO_CAN_DISABLE)
609 ec->clean_nlp = rx;
610
611 /* Output scaled back up again to match input scaling */
612
613 return (int16_t) ec->clean_nlp << 1;
614}
615EXPORT_SYMBOL_GPL(oslec_update);
616
617/* This function is separated from the echo canceller is it is usually called
618 as part of the tx process. See rx HP (DC blocking) filter above, it's
619 the same design.
620
621 Some soft phones send speech signals with a lot of low frequency
622 energy, e.g. down to 20Hz. This can make the hybrid non-linear
623 which causes the echo canceller to fall over. This filter can help
624 by removing any low frequency before it gets to the tx port of the
625 hybrid.
626
627 It can also help by removing and DC in the tx signal. DC is bad
628 for LMS algorithms.
629
630 This is one of the classic DC removal filters, adjusted to provide
631 sufficient bass rolloff to meet the above requirement to protect hybrids
632 from things that upset them. The difference between successive samples
633 produces a lousy HPF, and then a suitably placed pole flattens things out.
634 The final result is a nicely rolled off bass end. The filtering is
635 implemented with extended fractional precision, which noise shapes things,
636 giving very clean DC removal.
637*/
638
639int16_t oslec_hpf_tx(struct oslec_state *ec, int16_t tx)
640{
641 int tmp;
642 int tmp1;
643
644 if (ec->adaption_mode & ECHO_CAN_USE_TX_HPF) {
645 tmp = tx << 15;
646
647 /*
648 * Make sure the gain of the HPF is 1.0. The first can still
649 * saturate a little under impulse conditions, and it might
650 * roll to 32768 and need clipping on sustained peak level
651 * signals. However, the scale of such clipping is small, and
652 * the error due to any saturation should not markedly affect
653 * the downstream processing.
654 */
655 tmp -= (tmp >> 4);
656
657 ec->tx_1 += -(ec->tx_1 >> DC_LOG2BETA) + tmp - ec->tx_2;
658 tmp1 = ec->tx_1 >> 15;
659 if (tmp1 > 32767)
660 tmp1 = 32767;
661 if (tmp1 < -32767)
662 tmp1 = -32767;
663 tx = tmp1;
664 ec->tx_2 = tmp;
665 }
666
667 return tx;
668}
669EXPORT_SYMBOL_GPL(oslec_hpf_tx);
670
671MODULE_LICENSE("GPL");
672MODULE_AUTHOR("David Rowe");
673MODULE_DESCRIPTION("Open Source Line Echo Canceller");
674MODULE_VERSION("0.3.0");